Compressing Recurrent Neural Networks Using Hierarchical Tucker Tensor Decomposition